Scenario-aware dynamic graph learning with topology memory for wind farm cluster power forecasting

To address the strong heterogeneity of output scenarios and the pronounced time-varying correlation structures in wind farm clusters, this paper proposes a scenario-aware multi-channel time-frequency domain memory-enhanced graph convolutional network (MCTF-MEGCN) for wind power forecasting. First, a Composite Wind Intensity Index (CWI) is constructed from multiple feature dimensions to characterize the overall operating state of the wind farm cluster and realize differentiated scenario partitioning. Second, according to the scenario labels, different correlation calculation windows are selected to construct multi-time-scale dynamic graphs, enabling the dynamic characterization of evolving dependencies among wind farms. Furthermore, four types of correlation graphs are established in the time and frequency domains from spatial and temporal perspectives. Memory-enhanced graph convolution module (MEGCN) is introduced to integrate historical topology information with current dynamic graph information, and scenario-label embedding is used to guide the adaptive fusion of multi-channel features. Finally, case studies are conducted using data from two wind farm clusters located in Jilin and western Inner Mongolia. Compared with traditional forecasting baseline models, the proposed method achieves average relative reductions of 15.80% and 15.03% in NRMSE for the Jilin and western Inner Mongolia wind farm clusters, respectively. These results demonstrate that the proposed method achieves favorable forecasting accuracy and applicability across different regions, wind resource conditions, and spatial configurations of wind farms, thereby providing useful support for wind farm cluster grid integration and power-system dispatch decisions.